Universal Representation Learning from Multiple Domains for Few-shot Classification

ICCV 2021

Wei-Hong Li, Xialei Liu and Hakan Bilen

Abstract
In this paper, we look at the problem of few-shot image classification that aims to learn a classifier for previously unseen classes and domains from few labeled samples. Recent methods use various adaptation strategies for aligning their visual representations to new domains or select the relevant ones from multiple domain-specific feature extractors. In this work, we present URL, which learns a single set of universal visual representations by distilling knowledge of multiple domain-specific networks after co-aligning their features with the help of adapters and centered kernel alignment. We show that the universal representations can be further refined for previously unseen domains by an efficient adaptation step in a similar spirit to distance learning methods. We rigorously evaluate our model in the recent Meta-Dataset benchmark and demonstrate that it significantly outperforms the previous methods while being more efficient.
Universal Representation Learning

Our method (illustrated in Figure 1) learns a single universal feature extractor \(f_{\phi}\) that is distilled from multiple feature extractors \(\{f_{\phi^{\ast}_\tau}\}_{\tau}^{K}\) during meta-training.

Figure 1. URL - Universal Representation Learning.
Feature adaptation in meta-test

In meta-test stage, we use a linear transformation \(A_{\vartheta}\) to further refine the universal representations to unseen domains as illustrated in Figure 2.

Figure 2. Feature adaptation procedure in meta-test.
Results
We evaluate our method on Meta-dataset. Please refer to our paper and the leaderboard of Meta-dataset for more results.
Test Datasets URL (Ours) MDL Best SDL URT [6] SUR [4] Simple CNAPS [5] CNAPS [2] BOHB-E [3] Proto-MAML [1]
Avg rank 1.2 4.8 4.8 4.2 5.4 4.8 6.8 8.0 7.7
ImageNet 57.5±1.1  52.9±1.2  54.3±1.1  55.0±1.1  54.5±1.1  56.5±1.1  50.8±1.1  51.9±1.1  46.5±1.1 
Omniglot 94.5±0.4  93.7±0.5  93.8±0.5  93.3±0.5  93.0±0.5  91.9±0.6  91.7±0.5  67.6±1.2  82.7±1.0 
Aircraft 88.6±0.5  84.9±0.5  84.5±0.5  84.5±0.6  84.3±0.5  83.8±0.6  83.7±0.6  54.1±0.9  75.2±0.8 
Birds 80.5±0.7  79.2±0.8  70.6±0.9  75.8±0.8  70.4±1.1  76.1±0.9  73.6±0.9  70.7±0.9  69.9±1.0 
Textures 76.2±0.7  70.9±0.8  72.1±0.7  70.6±0.7  70.5±0.7  70.0±0.8  59.5±0.7  68.3±0.8  68.2±0.8 
Quick Draw 81.9±0.6  81.7±0.6  82.6±0.6  82.1±0.6  81.6±0.6  78.3±0.7  74.7±0.8  50.3±1.0  66.8±0.9 
Fungi 68.8±0.9  63.2±1.1  65.9±1.0  63.7±1.0  65.0±1.0  49.1±1.2  50.2±1.1  41.4±1.1  42.0±1.2 
VGG Flower 92.1±0.5  88.7±0.6  86.7±0.6  88.3±0.6  82.2±0.8  91.3±0.6  88.9±0.5  87.3±0.6  88.7±0.7 
Traffic Sign 63.3±1.2  49.2±1.0  47.1±1.1  50.1±1.1  49.8±1.1  59.2±1.0  56.5±1.1  51.8±1.0  52.4±1.1 
MSCOCO 54.0±1.0  47.3±1.1  49.7±1.0  48.9±1.1  49.4±1.1  42.4±1.1  39.4±1.0  48.0±1.0  41.7±1.1 
MNIST 94.5±0.5  94.2±0.4  91.0±0.5  90.5±0.4  94.9±0.4  94.3±0.4 
CIFAR-10 71.9±0.7  63.2±0.8  65.4±0.8  65.1±0.8  64.2±0.9  72.0±0.8 
CIFAR-100 62.6±1.0  54.7±1.1  56.2±1.0  57.2±1.0  57.1±1.1  60.9±1.1 

[1] Eleni Triantafillou, Tyler Zhu, Vincent Dumoulin, Pascal Lamblin, Utku Evci, Kelvin Xu, Ross Goroshin, Carles Gelada, Kevin Swersky, Pierre-Antoine Manzagol, Hugo Larochelle; Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples; ICLR 2020.

[2] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, Richard E. Turner; Fast and Flexible Multi-Task Classification Using Conditional Neural Adaptive Processes; NeurIPS 2019.

[3] Tonmoy Saikia, Thomas Brox, Cordelia Schmid; Optimized Generic Feature Learning for Few-shot Classification across Domains; arXiv 2020.

[4] Nikita Dvornik, Cordelia Schmid, Julien Mairal; Selecting Relevant Features from a Multi-domain Representation for Few-shot Classification; ECCV 2020.

[5] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, Leonid Sigal; Improved Few-Shot Visual Classification; CVPR 2020.

[6] Lu Liu, William Hamilton, Guodong Long, Jing Jiang, Hugo Larochelle; Universal Representation Transformer Layer for Few-Shot Image Classification; ICLR 2021.

Qualitative Results

We also qualitatively analyze our method and compare it to URT [6] in Figure 3 by illustrating the nearest neighbors in four different datasets given a query image (see supplementary for more examples). While URT retrieves images with more similar colors, shapes and backgrounds, our method is able to retrieve semantically similar images and finds more correct neighbors than URT. It again suggests that our method is able to learn more semantically meaningful and general representations.

Figure 3. Qualitative analysis of our method in four datasets.

Updated: